This page lists the per-million-token input and output rates that Anthropic, OpenAI, Google, and xAI publish for their current flagship and mid-tier models. The mid-tier band has converged into a narrow range, with $2 per million input tokens at Anthropic, OpenAI and Google and $1.25 at xAI. Below that sits a cheap band that three of the four vendors price under $1 per million input tokens, and every vendor publishes a separate rate once a prompt passes a long-context threshold - though the thresholds themselves differ, running from 100K tokens at Anthropic to 272K at OpenAI.
Verified 20 November 2026 against each vendor’s current pricing page.
What changed since this article was first drafted: headline Anthropic pricing moved from Claude Opus 5 ($5/$25) and Claude Sonnet 5 ($2/$10) to Claude Opus 5.5 ($4/$20) and Claude Sonnet 5.5 ($2/$10); Claude Haiku 5.5 introduced a 100K-context tier at $0.10/$0.50 with a higher rate above that threshold, and Claude Fable 5.2 now leads Anthropic’s card at $8/$40, with Claude Fable 5.1 demoted to the legacy list at $10/$50. OpenAI added the gpt-6 family at $10/$50 (gpt-6-astra), $2/$10 (gpt-6.1-sol, gpt-6-sol), and $0.10/$0.50 (gpt-6-luna), then gpt-6.2-astra at $8/$40 and gpt-6.2-sol at $2/$10. Google Gemini 3.1 Pro Preview carries a separate 200K threshold ($2/$12 short context, $4/$18 long context), and the Gemini 4 family arrived as four paid models - Gemini 4 Neon at $3/$15, Gemini 4 Carbon and Gemini 4 Argon at $2/$10 doubling to $4/$20 on 1 December 2026, and Gemini 4 Hydrogen at $1.50/$7.50. Gemini 3.8 Flash rates are scheduled to roughly double on 1 January 2027. xAI Grok 4.5 through 4.7 sit at $2/$6 below 200K and $4/$12 at or above. Google has also deprecated Gemini 2.5 Pro and Gemini 2.5 Flash: from 1 December 2026, requests to either model automatically route to Gemini 3.8 Flash and Gemini 3.5 Flash-Lite respectively, and are billed at those models’ rates.
The Headline Tier
The flagship or top-tier rate card is where vendors most aggressively signal where they think the price ceiling sits today. The comparison table below lists the current top-tier or near-top-tier model from each vendor and the rates each one publishes as of 20 November 2026.
| Vendor | Current flagship | Input $ / MTok | Output $ / MTok | Long-context tier (threshold varies by vendor) |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.2 | 8.00 | 40.00 | none published |
| Anthropic | Claude Opus 5.5 | 4.00 | 20.00 | none published |
| OpenAI | gpt-6.2-astra | 8.00 | 40.00 | yes - $16 / $60 above 272K tokens |
| OpenAI | gpt-6-astra | 10.00 | 50.00 | yes - $20 / $75 above 272K tokens |
| OpenAI | gpt-5.5-pro | 30.00 | 180.00 | yes - $60 / $270 above 272K tokens |
| Gemini 4 Neon | 3.00 | 15.00 | not in published rate card | |
| Gemini 3.1 Pro Preview | 2.00 (≤200K) | 12.00 (≤200K) | yes - $4 / $18 above 200K tokens | |
| Gemini 4 Carbon and Argon (paid) | 2.00 (then 4.00 from 1 Dec 2026) | 10.00 (then 20.00 from 1 Dec 2026) | not in published rate card | |
| xAI | grok-4.5 / 4.6 / 4.7 | 2.00 (<200K) | 6.00 (<200K) | yes - $4 / $12 above 200K tokens |
Three patterns show up. First, xAI’s Grok 4.5 through 4.7 publish the same input and output rate, so the price ceiling at the top is no longer a competition move within that vendor - the question is which generation pair, at the same rate. Their cached-input rates do still differ: $0.30 on grok-4.5 against $0.50 on grok-4.6 and grok-4.7. Second, the Anthropic and OpenAI flagships now meet at the same rate: Claude Fable 5.2 and gpt-6.2-astra both sit at $8/$40, which suggests the two vendors have stopped racing each other at the very top and are focused on different competitors. The older Claude Fable 5.1 stays listed at $10/$50, the same rate as gpt-6-astra (Anthropic pricing, Anthropic API pricing, OpenAI pricing). Third, Google’s own ceiling sits below both of them: Gemini 4 Neon leads at $3/$15, with Gemini 4 Carbon and Argon at $2/$10 and Gemini 4 Hydrogen at $1.50/$7.50 (Gemini pricing).
The Mid-Tier Band Is Where the Race Lives
The mid-tier is the rate most readers actually pay when they integrate a hosted model. Here the bands have converged near $2 input / $10 output at Anthropic, OpenAI and Google, with xAI’s mid-tier sitting lower at $1.25 input and $2.50 output.
| Vendor | Mid-tier model | Input $ / MTok | Output $ / MTok | Notes |
|---|---|---|---|---|
| Anthropic | Claude Sonnet 5.5 | 2.00 | 10.00 | cache reads $0.10 / MTok |
| OpenAI | gpt-6-sol | 2.00 | 10.00 | long context $4 / $15 above 272K |
| OpenAI | gpt-6.1-sol | 2.00 | 10.00 | cached input $0.10, long context $4 / $15 above 272K |
| OpenAI | gpt-6.2-sol | 2.00 | 10.00 | cached input $0.10, long context $4 / $15 above 272K |
| OpenAI | gpt-5.6-sol | 4.00 | 20.00 | promotional until 30 November 2026 |
| OpenAI | gpt-5 | 1.25 | 10.00 | older generation, still listed |
| Gemini 2.5 Pro | 1.25 (≤200K) | 10.00 (≤200K) | long context $2.50 / $15 above 200K; deprecated - from 1 Dec 2026 requests route to Gemini 3.8 Flash at its rates | |
| Gemini 3.1 Pro Preview | 2.00 (≤200K) | 12.00 (≤200K) | long context $4 / $18 above 200K | |
| Gemini 4 Hydrogen | 1.50 | 7.50 | 1M-token context, no published step-up | |
| xAI | grok-4.3 | 1.25 (<200K) | 2.50 (<200K) | long context $2.50 / $5.00 above 200K |
Both Anthropic and OpenAI publish cache-read prices that matter at this tier. Anthropic lists cache reads at $0.10 per million tokens for Sonnet 5.5 - one twentieth of the $2.00 input rate - and that means any workload that reuses a system prompt cuts the input bill by an order of magnitude before it runs. OpenAI publishes the same discount on its own tables: gpt-6.2-astra lists cached input at $0.40 against an $8.00 input rate, gpt-6.1-sol at $0.10 against a $2.00 input rate, gpt-6-astra at $1.00 against $10.00, gpt-5-mini at $0.025 against $0.25, and gpt-4.1-mini at $0.10 against $0.40. Google publishes its caching prices in a separate “Context caching price” line rather than a cached-input column, and adds an hourly storage charge on top: Gemini 3.1 Pro Preview lists $0.20 at or below 200K against a $2.00 input rate, Gemini 2.5 Pro $0.125 against $1.25, Gemini 2.5 Flash $0.03 against $0.30, Gemini 2.5 Flash-Lite $0.01 against $0.10, and Gemini 4 Carbon and Argon $0.10 against $2.00 (Gemini pricing).
The local-ai companion to this comparison is the crossover article, which uses these exact rates on the API side and weighs them against the electricity-plus-hardware cost of running an open-weight model on a consumer GPU.
Below the Mid-Tier: The Cheap Band
The cheap band sits below the mid-tier, and three of the four vendors publish a sub-$1 input rate here.
| Vendor | Cheap-tier model | Input $ / MTok | Output $ / MTok |
|---|---|---|---|
| Anthropic | Claude Haiku 5.5 (≤100K context) | 0.10 | 0.50 |
| Anthropic | Claude Haiku 5.5 (>100K context) | 0.50 | 2.50 |
| OpenAI | gpt-6-luna | 0.10 | 0.50 |
| OpenAI | gpt-5-nano | 0.05 | 0.40 |
| OpenAI | gpt-5-mini | 0.25 | 2.00 |
| OpenAI | gpt-4.1-mini | 0.40 | 1.60 |
| Gemini 2.5 Flash-Lite | 0.10 | 0.40 | |
| Gemini 2.5 Flash (deprecated - from 1 Dec 2026 requests route to Gemini 3.5 Flash-Lite at its rates) | 0.30 | 2.50 | |
| Gemini 3.5 Flash-Lite | 0.30 | 2.50 | |
| Gemini 3.6 Flash | 0.75 | 3.75 (then $1.50 / $7.50 from 1 Jan 2027) | |
| Gemini 3.8 Flash | 0.75 | 3.75 (then $1.50 / $7.50 from 1 Jan 2027) | |
| xAI | grok-build-0.1 | 1.00 (<200K) | 2.00 (<200K) - doubles above 200K, see below |
The catch on the cheap band is Google Gemini 3.8 Flash’s published increase. The Gemini rate card lists Gemini 3.8 Flash at $0.75 input / $3.75 output now, with an explicit note that those rates double to $1.50 input / $7.50 output on 1 January 2027 (Gemini pricing). Any reader who picks the cheap tier expecting today’s rate should also re-check on that date. Gemini 3.6 Flash carries the same future-increase pattern on the same dates.
The Verbs: Batch, Flex, Priority, Fast, Ultrafast
Every vendor names their service tiers differently. The four that matter for cost engineering are the same idea under different names: the cheaper tier for asynchronous work, and the more expensive tier for low-latency work.
| Vendor | Standard | Cheaper async tier | Rate | Premium latency tier | Rate |
|---|---|---|---|---|---|
| Anthropic | standard | Batch | 50% off | Fast (Fable 5.2 and Opus 5.5) | 2x standard pricing |
| OpenAI | standard | Batch | half of standard | Fast (formerly Priority) | 2x on the gpt-6 family and most gpt-5 models; 2.5x on gpt-5.5; 1.8x on gpt-5-mini, 1.75x on gpt-4.1-mini, 1.7x on gpt-4o and 1.67x on gpt-4o-mini; no Fast rate published for gpt-5.5-pro |
| OpenAI | standard | Flex | half of standard | Ultrafast | 6x on gpt-6-astra and gpt-6.1-sol; 3x on gpt-6.2-astra and gpt-6.2-sol through 28 February 2027, then 6x |
| standard | Batch / Flex | 50% off (Batch); Flex listed at the same rates | Priority | 1.8x, no percentage published | |
| xAI | standard | no rates published | - | no rates published | - |
The Batch discount is half of standard across Anthropic, OpenAI and Google, though they publish it differently: Anthropic writes “Save 50% with batch processing”, Google writes “Batch API (50% cost reduction)”, and OpenAI lists Batch and Flex prices at half the standard rate for most models. Google’s rate card names the latency tier Priority and publishes Priority prices without a percentage; those prices are 1.8x standard, with Gemini 4 Argon at $3.60 against a $2.00 standard input rate. Anthropic’s fast mode covers Fable 5.2 and Opus 5.5 at 2x standard pricing, and OpenAI’s Fast tier holds at 2x on the gpt-6.2 models as well, listing $16/$80 against a standard $8/$40 for gpt-6.2-astra. OpenAI’s Fast and Ultrafast tables also list prices rather than multipliers: Fast lists $20/$100 against a standard $10/$50 for gpt-6-astra, and Ultrafast lists $60/$300 for the same model, six times the standard rate. Flex is being wound down: the page says it “remains available on the models listed here as we wind down the beta”, the newest models are not on it, and Ultrafast on gpt-6.2-astra and gpt-6.2-sol is priced at 3x standard through 28 February 2027 and 6x from 1 March 2027. xAI’s models page notes that not every model accepts Batch API requests, but publishes no Batch rates (xAI Grok pricing). Where a workload tolerates asynchronous turnaround, Batch is the cheapest rate on the rate card; whether that beats running the model on your own hardware is the question the cost crossover article works through.
Long-Context Step-Ups
All four vendors publish a separate rate for prompts above a context threshold, though not for every model they sell. The threshold itself is not standardised: it runs from 100K tokens at Anthropic to 272K at OpenAI, with Google and xAI both using 200K.
| Vendor | Threshold above which the long-context rate kicks in | Effect on input / output |
|---|---|---|
| Anthropic | 100K tokens for Haiku 5.5 only | $0.10 -> $0.50 input, $0.50 -> $2.50 output |
| OpenAI | 272K input tokens (gpt-6.2 / gpt-6.1 / gpt-6 / gpt-5.6 family and above) | roughly 2x the standard input rate, 1.5x the output |
| 200K tokens (Gemini 3.1 Pro Preview, Gemini 2.5 Pro - deprecated) | $2 -> $4 input, $12 -> $18 output at Pro Preview | |
| xAI | 200K tokens (Grok 4.3 through 4.7) | $2 -> $4 input, $6 -> $12 output |
| xAI | 200K tokens (grok-build-0.1) | $1 -> $2 input, $2 -> $4 output |
Important catch: xAI’s rate card states that when a request reaches the threshold, the whole request is billed at the higher rate, not just the tokens above it - “requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request”. A 201K-token prompt does not get cheaper than a 200K-token prompt. Anthropic prices the full 1M-token window at standard rates on Claude 4.6 and later, which leaves Haiku 5.5 as its only prompt-length tier, and Google’s Gemini 4 models take a 1M-token window at standard rates, so Google’s published step-ups cover Gemini 3.1 Pro Preview and the deprecated Gemini 2.5 Pro only (Anthropic API pricing, Gemini pricing). Anthropic, OpenAI and Google publish the threshold and the higher rate but do not spell the billing rule out on the pages cited here.
Bottom Line
The headline numbers tell one story; the verbs and the thresholds tell the second one. The mid-tier band has converged into a $1.25 to $2 input range, and the cheap tier sits below $1 per million input tokens at three of the four vendors. A reader who lives at the cheap tier should treat the Gemini 3.8 Flash rate-card note (the 1 January 2027 doubling) as a real re-pricing event and not as marketing - it is the published rate, not a rumored change.
For anyone considering whether to pay these rates or run a model at home, the cost crossover article walks the same hosted prices against consumer-GPU hardware costs. For a hub listing the model capabilities rather than the price points, the leading-frontier-models comparison is the companion page.
This page tracks the four major vendors’ published rates. Each vendor updates its own page without notice; if a reader makes a buying decision on these numbers, the step before spending is to re-open the cited pricing page and re-read the row they are about to commit to.